Quasar 1.1 438B is the first documented model to use quantum-generated data in its training, produced on a 156-qubit IBM Heron r2 processor at IBM Quantum System Two in Donostia-San Sebastián (Spain). The work came out of the Research division at Multiverse Computing, and it is an experiment: it shows what our teams can build at the intersection of quantum computing and language models, not a shortcut to a better model. The gains in the release notes come from broader healing data, verbosity tuning and refusal steering. Here is how the quantum side of it was done.
The encoding problem
To put any part of a language model's computation on a quantum device, you first have to get its hidden state into a qubit register. Qwen3-30B-A3B carries a hidden dimension of 2048, and the two standard encodings both fail at that width, in opposite directions. Amplitude encoding packs D amplitudes into log₂ D qubits, but loading a generic vector requires circuit depth O(D), beyond what NISQ hardware can execute coherently. Angle encoding needs one shallow layer of gates, but its width grows linearly with the number of features.
Our route keeps angle encoding and makes each qubit carry more of the load. Dense angle encoding writes two features into every qubit, as an RY and an RZ rotation, and layerwise uploads then send different features to the same qubits at successive layers instead of re-uploading the same ones. Eight qubits, two features each, four layers: 64 features per register, without adding a single qubit. What it costs is the extra Fourier frequencies that re-uploading would have supplied, traded for multivariate interactions among the encoded features in the measured Pauli expectations.
Multi-head quantum neural networks
That encoding is what makes the multi-head QNN possible. The 2048-dimensional hidden space is partitioned into k = 32 disjoint heads, each angle-encoded into its own 8-qubit register, processed in parallel, and recoupled through sparse classical linear layers we call the mixers. The output returns to the residual stream scaled by a learned factor α. The mixers are deliberately constrained: an unconstrained dense matrix would be expressive enough to carry the task alone, reducing the quantum module to an ornament. Stripping the quantum heads out of the trained block and leaving that scaffolding in place costs 3.5 points of perplexity, so the circuits are doing the work.
Angles are trainable affine functions a_i x_i + b_i, so each head learns its own Fourier scales. Entanglement comes from an open chain of seven RZZ gates. Readout uses a pool of 64 Pauli observables per head.
One such block, 145,281 trainable parameters, replaces a contiguous span of eight transformer blocks in the 48-layer model, displacing 16.33% of its parameters. Each displaced block carries 6.23 × 10⁸ parameters of its own.
On the device
The trained circuits were executed on ibm_basquecountry, a 156-qubit IBM Heron r2 processor at IBM Quantum System Two in Donostia-San Sebastián, with nothing retrained or rescaled for the step. The 32 heads of a token do not fit at once; they are covered by three circuits in a 14/14/4 split, the largest compiling to 392 two-qubit gates at a two-qubit depth of 13 with zero routing swaps. Compiled cost is constant in the size of the absorbed span: a 16-layer replacement costs the device what a single-layer one does.
From hybrid model to healing data
The generator is a Qwen3-30B-A3B running on GPU with eight of its layers replaced by the multi-head QNN block. Two stages train it: a local one by gradient descent, where the module learns to reproduce the output of the span it replaces, and a global one combining supervised fine-tuning with knowledge distillation. The prompts are real text from a web mixture: FineWeb-Edu 43%, Wikipedia 25%, DCLM 23%, the rest mathematics and Cosmopedia.
Generation then ran in two regimes. In the first, tokens are routed to the quantum processor itself. Each token takes around a minute and a half on the device, so a sentence of roughly 15 tokens takes about 25 minutes. That regime puts real hardware in the generation path, and at that rate it yields a proof of sentences.
The volume came from the second regime, an exact statevector simulation of the same circuits, GPU-based, with noise applied from calibration parameters measured on the IBM device. That produced 16,500 samples, around 4.2 million tokens, in a few hours.
The quantum processor’s contribution is therefore small in volume and larger in what it establishes: the path from a quantum device to a production training set exists and has been walked end to end. Widening it is the work ahead.
Where it stops
The multi-head QNN is a position-wise map, but the block it replaces contains self-attention, so every replacement deletes cross-token mixing at that depth. Throughput is the other hard limit, and the 25 minutes per sentence above is what it looks like in practice. Circuit depth remains open: better ansätze, native gate sets and error mitigation all have room there.
What's new in Quasar 1.1
The quantum-inspired tensor network methods behind CompactifAI’s expert selection came out of this same line of work. What is new in Quasar 1.1 is that the quantum side is no longer only an inspiration for classical algorithms: part of the data that shaped the model came off a quantum processor in the city where we build these models. It is experimental and it makes the model no better than a classical route would have. A first step in a direction is the whole claim.